Glossary
Storage performance
Storage performance is how quickly a storage system completes the work applications send it, described by three linked measures: latency, IOPS and throughput. No single figure captures it, because each measure answers a different question and a system can lead on one while trailing on another.
Three measures, three questions
| Measure | Expressed as | Question it answers | Workloads that depend on it |
|---|---|---|---|
| Latency | Milliseconds or microseconds per request | How long does one request wait? | Databases, virtual machines, interactive applications |
| IOPS | Operations completed per second | How many separate requests finish each second? | Transactional systems, small files, metadata lookups |
| Throughput | Megabytes or gigabytes per second | How much data moves each second? | Backup and restore, analytics scans, media, training data reads |
A buyer comparing systems needs all three, measured on the same workload. A datasheet that leads with one and omits the others usually picked the one that flatters the product.
How the measures pull against each other
The three are linked, so improving one often costs another. Request size connects IOPS and throughput: a system reports its highest operation count with tiny requests and its highest data rate with large ones, and rarely reaches both in the same test. Load connects IOPS and latency: pushing more operations through lengthens the queue, so each request waits longer, and an operation rate quoted without the latency at that rate describes a point nobody would run in production. Access pattern affects all three, which is the subject of sequential vs random I/O.
Why results change after day one
A new system is empty, its caches hold most of what is read, its flash has plenty of free space and nothing is being repaired. Each of those conditions fades. Data growth pushes the working set out of cache, so more reads reach slower media. Flash with little free space takes longer to accept writes. More applications share the same drives and links. Background work, such as rebuilding after a drive failure, checking data integrity and moving data between tiers, competes with production traffic. The hardware is unchanged, yet the same workload runs slower in year two than in the acceptance test. Acceptance tests run near the expected fill level, with a rebuild in progress, give a far better forecast.
Where do benchmarking, tuning and QoS fit?
They act on performance at different moments. Benchmarking measures it under controlled conditions so systems and configurations can be compared. Tuning uses measurements on one system to find and remove whatever is limiting a workload. Storage QoS decides how a shared system's performance is split between workloads once it is in service. Hardware layers such as a flash cache or persistent memory change latency for particular kinds of requests rather than for the system as a whole.
Frequently asked questions
How is storage performance measured?
With latency, IOPS and throughput reported together, including the slowest percentiles of latency, on a workload that resembles production in request size, read and write mix and concurrency.
Does adding capacity add performance?
Only when the new capacity brings its own drives, processors and network ports. Replacing drives with larger ones in the same servers adds space without adding speed.